Reproduce “S-NeRF: Neural Radiance Fields for Street Views” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Ziyang Xie, Junge Zhang, Wenye Li, Feihu Zhang, Li Zhang
S-NeRF, a Neural Radiance Fields framework, generates photorealistic novel views for both vast street scenes and dynamic vehicles using standard self-driving datasets. The method achieved up to 45% higher PSNR and 26% higher SSIM for background scenes and 45% higher PSNR for moving vehicles compared to baselines on nuScenes and Waymo datasets.
Key insights
- Transforming camera poses relative to a dynamic object effectively simplifies its reconstruction, treating the moving object as static within a new coordinate system.
- Robust novel view synthesis in street scenes necessitates specialized scene parameterization for unbounded backgrounds and a confidence-guided approach to leverage noisy, sparse LiDAR depth.
- A learnable confidence mechanism, combining photometric, feature, and geometric consistencies, can effectively mitigate the impact of outliers in depth supervision from real-world self-driving datasets.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.